I forked llama.cpp and added some crude mechanism to keep exactly this behavior under control - essentially guiding the reasoning process by injecting text strategically at specific thresholds. This was mainly put together to rein in Qwen3.6-27B, but I'd imagine 3.8 would react similarly.
Of course hacks like this are not perfect and may degrade performance slightly due to injected text pushing the model slightly out-of-distribution, so the string constants need to be chosen carefully - Qwen3.5's technical whitepaper does provide some guidance in this regard. The mechanism is absolutely more of a hack than a feature, and i'd imagine will be made redundant once llama.cpp supports more appropriate reasoning controls - but for now, i've found it pretty useful.
I did make a PR to the official llama-cpp repo some time back (about a month or so), but abandoned it as there seemed to be too much community concern that the mechanism would degrade model performance... Perhaps i'll polish it up and put some effort into benchmarking and revive the project in the near future.
This will degrade performance significantly. LLama.cpp has had this for a while and it tanks benchmark performance. I ran GPQA on GLM 5.2 using the llama implementation and it came back 19 points under the regular results.
I'm Laurence - a senior engineer and manager with 10 years' experience in full-stack development roles, with significant experience as a FDE - as well as team management and advisory roles - extensive experience working with multinational clients in the financial services and insurance industries.
4.5 years' experience in AI (Conversational/Voice as well as advanced RAG systems development and implementation). Additional experience in embedded systems development and sensor fusion.
I'd love to hear what's out there - so please do feel free to reach out if you'd like a full copy of my resume or have a chat! :)
I'm always curious on how security software can provide a ROI.
I had McAfee tell me one time that the hackersafe logo on our website would increase sales by 10%, this was at a Fortune 50 doing billions in sales online every year.
I was pretty hyped because it would have done wonders for my career, but then they walked it back and wouldn't explain it to me. I wasn't mad, I was disappointed.
I ran an AB test on 2012 not sure its relevant now, we tested the McAfee logo and conversion was boosted by 2%. Bigger boost was a lock icon, 3%. It kept increasing the more locks we added an topped at 5% after 5 lock icons.
Fork can be found here - https://github.com/laurencehardman/llama-mindcontrol/tree/ma...
Of course hacks like this are not perfect and may degrade performance slightly due to injected text pushing the model slightly out-of-distribution, so the string constants need to be chosen carefully - Qwen3.5's technical whitepaper does provide some guidance in this regard. The mechanism is absolutely more of a hack than a feature, and i'd imagine will be made redundant once llama.cpp supports more appropriate reasoning controls - but for now, i've found it pretty useful.